Monaural speech filter
A system receives monaural sound which includes speech and background noises. The received sound is divided by frequency and time into time-frequency units (TFUs). Each TFU is classified as speech or non-speech by a processing unit. The processing unit for each frequency range includes at least one of a deep neural network (DNN) or a linear support vector machine (LSVM). The DNN extracts and classifies the features of the TFU and includes a pre-trained stack of Restricted Boltzmann Machines (RBM), and each RBM includes a visible and a hidden layer. The LSVM classifies each TFU based on extracted features from the DNN, including those from the visible layer of the first RBM, and those from the hidden layer of the last RBM in the stack. The LSVM and DNN include training with a plurality of training noises. Each TFU classified as speech is output.
1. A system separating speech from background noises, comprising:
a cochlear filter which receives monaural sound in an electronic format which includes speech and a plurality of background noises and divides the received sound by frequency and time into a plurality of time-frequency units;
one or more processing units which classify each time-frequency unit as speech or non-speech and output each time-frequency unit classified as speech, the processing unit for each frequency range including at least one of:
a deep neural network which extracts and classifies the features of the time-frequency unit and includes a pre-trained stack of Restricted Boltzmann Machines, each Restricted Boltzmann Machine includes a visible layer and a hidden layer; or
a linear support vector machine which classifies the each time-frequency unit based on extracted features from the deep neural network and includes visible features and hidden features, the visible features include a first visible layer of a first Restricted Boltzmann Machine in the pre-trained stack of Restricted Boltzmann Machines, and the hidden features includes a hidden layer of a last Restricted Boltzmann Machine in the stack, and
wherein the linear support vector machine and deep neural network include training with a plurality of training noises.
2. The system according to claim 1 , wherein the plurality of training noises exceeds 10 distinct background noises.
3. The system according to claim 1 , further including:
a sound receiving device which converts the monaural sound waves to an electronic format.
4. The system according to claim 1 , further including:
a sound synthesizer which combines the output time-frequency units.
5. The system according to claim 4 , wherein the sound synthesizer outputs the combined output as monaural sound waves.
6. The system according to claim 1 , wherein the frequency includes the range 50 Hz to 8000 Hz.
7. The system according to claim 1 , wherein the time of each time-frequency unit includes 20 milliseconds shifted by 10 milliseconds.
8. The system according to claim 1 , wherein the number of frequency ranges includes 1-64, and each frequency range is cross correlated.
9. A method of separating monaural speech from background noise, comprising:
receiving monaural sound in an electronic format which includes speech and a plurality of background noises;
dividing the received sound by frequency into a plurality of channels, and the sound of each channel by time into a plurality of time-frequency units;
classifying each time-frequency unit as speech or non-speech by a processing unit, wherein the processing unit for each channel includes at least one of:
a deep neural network which extracts and classifies the features of the time-frequency unit and includes a pre-trained stack of Restricted Boltzmann Machines, each Restricted Boltzmann Machine includes a visible layer and a hidden layer; or
a trained linear support vector machine which classifies each time-frequency unit based on extracted features from the deep neural network and includes visible features and hidden features, the visible features include a first visible layer of a first Restricted Boltzmann Machine in the pre-trained stack of Restricted Boltzmann Machines, and the hidden features include a hidden layer of a last Restricted Boltzmann Machine in the stack; and
wherein the linear support vector machine and deep neural network include training with a plurality of training noises; and
outputting each time-frequency unit classified as speech.
10. The method according to claim 9 , wherein the plurality of training noises exceeds 10 distinct background noises.
11. The method according to claim 9 , further including:
converting the monaural sound waves to an electronic format.
12. The method according to claim 9 , wherein outputting further includes:
combining the output time-frequency units.
13. The method according to claim 9 , wherein outputting further includes outputting the combined output as monaural sound waves.
14. The method according to claim 9 , wherein the frequency includes the range 50 Hz to 8000 Hz.
15. The method according to claim 10 , wherein the time of each time-frequency unit includes 20 milliseconds shifted by 10 milliseconds.
16. A system separating speech from background noises, comprising:
one or more processors programmed to:
receive a plurality of time-frequency units;
classify each time-frequency unit as speech or non-speech; and
output each time-frequency unit classified as speech;
wherein each processor is programmed for a corresponding frequency range, the processor for each frequency range including:
a deep neural network which extracts and classifies the features of the time-frequency unit and includes a pre-trained stack of Restricted Boltzmann Machines, each Restricted Boltzmann Machine includes a visible layer and a hidden layer; and
a linear support vector machine which classifies each time-frequency unit based on extracted features from the deep neural network and includes visible features and hidden features, the visible features include a first visible layer of a first Restricted Boltzmann Machine in the pre-trained stack of Restricted Boltzmann Machines, and the hidden features includes a hidden layer of a last Restricted Boltzmann Machine in the stack, and
wherein the deep neural network is trained with raw acoustic features of the time-frequency units to generate learned features, and the linear support vector machine is trained with the learned features concatenated with the raw acoustic features; and
wherein the linear support vector machine and deep neural network include training with a plurality of training noises.
17. The system according to claim 16 , wherein the plurality of training noises exceeds 10 distinct background noises.
18. The system according to claim 16 , further including:
a cochlear filter which receives monaural sound in an electronic format which includes speech and a plurality of background noises and divides the received sound by frequency and time into the plurality of time-frequency units, the cochlear filter being configured to transmit the plurality of time-frequency units to the one or more processors.
19. The system according to claim 16 , further including:
a sound receiver which converts the monaural sound waves to an electronic format.
20. The system according to claim 16 , further including:
a sound synthesizer which combines the output time-frequency units as monaural sound waves.